Zero-Egress Health AI with a Patient-Held Consent Layer: Design and Early Operation of the LifeCare Network

Abstract

Personal health data is fragmented across device vendors, laboratories, clinics and apps, each holding it under its own terms, while a new class of AI assistants seeks to become the durable holder of an individual’s longitudinal health context. We describe the LifeCare Network, a de- ployed system that separates three concerns the industry usually collapses into one: control of health data (a patient-held identity, consent and authorization layer that stores no health data), holding of the data (provider-operated nodes that keep raw values in place), and inference over it (open-weight models that run on the node, with only interpreted outputs leaving). We present the consent receipt shaped to ISO/IEC TS 27560 and approved by a passkey assertion over the canonical grant, a UMA-pattern authorization server with purpose-bound tokens validated by introspection and revoked in under a second, a hash-chained per-person ledger and resource reg- istry, three enforcement layers that make “raw data never leaves” a property of code paths rather than policy, erasure by key destruction that preserves the ledger’s verifiability, and a software- rooted attestation whose measurement anyone can recompute from the node’s public response. We report early operating results from a single production node: a champion/challenger gate that rejected a stronger benchmark model on the network’s own golden set, a conversational layer redesigned after prompt-only rules failed in live traffic, and a pre-registered twelve-week outcome study with locked endpoints and kill criteria. We state the limits: attestation is not yet hardware-rooted, enforcement of caller authentication is a deployment setting, and no outcome data exists yet. We position the design against personal data stores, patient-controlled authorization servers, confidential cloud inference, and agent-memory layers, and argue that the unoccupied position is data-scoped, purpose-bound consent enforced next to the data.